Error band from VisualizeError

[Shilpi]

Hi,

I am fitting a distribution with Bernstein polynomial of order 1 using RooFIT.
I draw the error using VisualizeError(*r,1) where is RooFitResult pointer.

At some x (~140) value, I see that the error is ~0 [see the attached plot].

This is the error from linear
propagation so I did not expect this to go 0.

How do I understand this?

Thanks,
Shilpi

[amadio]

You have a larger error on the a coefficient of ax + b, and a smaller error on the b.
This could be due to a problem with your fit (like fixing b, for example, or not waiting for convergence).
Just looking at it, your line should probably have been further up from the plot you show.
Could you please paste here the parameters and their respective errors?

[Shilpi]

Thanks for the reply.

(1) The plot is set to have SetMinimum(0.001). There are zero entries which pull the fit down. So we have confirmed that this is fine.

(2) you are right about fixing the parameter. b is fixed. RooFIT says the status is 'OK’
And the error on a is large:

Value Error
6.59619e-01 2.24895e-01

In this case, its still not clear how the error band is ~0.
May be we could try to make B floating as well and then see if the error band around 140 increases?

Thanks much,
Shilpi

[amadio]

Yes, you need to let b vary, or the line you get will be far off from the points, as you can see in your plot.

[Shilpi]

I don’t think the fit being low is related to b not varied. As I mentioned that there are entries
with 0 events in several bins - not seen because the minimum is set to 0.01.
It can be seen in the attachment that even if we use 2 parameter fit, the fit is still lower
which is okay because of zero entry bins (not visible here).

I dont understand that if the error on the slope is large, constant being fixed, how can the error become ~0. I understand that the two lines Ax+B and (A+\deltaA)x + B have to intersect to make this happen and their intersection is not possible.

thanks again,

[Shilpi]

Here is the plot with SetMin(0) - same data events but with more fit functions.

[Shilpi]

My previous comment about intersection is not correct. They would intersect at x=0 in
ax+b parameterization. I will check how in RooBernstein the x is translated.